The situation
A US healthcare analytics provider needed to identify high-risk patients earlier and make sense of large volumes of unstructured clinical notes. The information was there, but spread across JSON records, SQL databases and free text that no model could use directly.
What we built
- Clinical NLP with spaCy, named entity recognition and rule-based extraction, processing around 500,000 patient records.
- ICD-10 and HCC code mapping using approximate matching against controlled vocabularies, extended with language models for automated classification of diagnostic and risk adjustment codes.
- Predictive models for annual claims cost and 30-day heart failure readmission, alongside disease classification.
- Pipelines unifying JSON, SQL and unstructured repositories into one governed layer, designed with HIPAA obligations in mind.
Results
- 500,000 patient records structured by the NLP pipeline
- 80 percent claim cost accuracy for small-bill customers
- 70 percent AUC on 30-day heart failure readmission risk
Why it matters
Most of the value in clinical data sits in free text that no model could use directly. Structuring it, and mapping it to the codes the business already works with, is what made risk and cost models possible on top of it.